Serving Mix

serving_mix creates mixed execution pressure representative of requests with different prompt and generation demands.

pantheon --test serving_mix --duration 60 --gpu 0 --mem 50 --verify --profile

Area	What it tests
Compute	Mixed projection-style work.
Memory	Synthetic request working-set reads.
Telemetry	Power, clocks, cache/DRAM behavior, and throttling.
Result	Synthetic requests per second.

This is not an HTTP server or queueing benchmark. Measure TTFT, inter-token latency, and production concurrency using the actual serving stack separately.